Mantis: Portable Toolkit for Building Secure Software
[!CAUTION] USE AT YOUR OWN RISK. BE EXTREMELY CAREFUL. This suite is
designed to generate and execute autonomously generated code that may be
unstable or perform unexpected actions. USE THIS ONLY IN ISOLATED,
RESTRICTED ENVIRONMENTS. Never run this suite on a machine with access to
production systems, sensitive data, or internal networks.
[!IMPORTANT] RESPONSIBLE USE AI models are non-deterministic and can
hallucinate findings or generate incorrect patches. All findings must be
manually verified by a security expert before being reported. Do not
mass-file unverified, AI-generated reports to open-source maintainers. A
failure to automatically reproduce a vulnerability does not definitively mean
it is a false positive, nor does a successful reproducer guarantee the bug is
exploitable in all contexts. Use Mantis responsibly.
[!WARNING] The pull request scan is untested. The GitHub Action in
action/, the mantis-pr-scan and mantis-knowledge-refresh workflows under
.github/workflows/, and reference/scripts/setup_github_action.sh have been
reviewed and exercised against stubs, but have not yet been run end-to-end
against a live project. Do not use the action to gate merges. The status note
at the top of
reference/docs/github-action.md lists what
execution has not verified and will be removed when it has.
Mantis is a set of skills along with an ADK reference harness for building
secure software in the new AI era of software development.
skills/ -- the method: the skills an agent loads on demand, from
threat modelling through research, review, patching and reporting.
reference/ -- an ADK harness that runs the method as a
pipeline, with its own two skills to configure and launch it. Meant to be
adapted; works as is.
action/ -- a GitHub Action that runs the harness over a pull
request (see the warning above).
Getting Started
Prerequisites
- OS: Linux, WSL2, or macOS (Apple Silicon supports local microVM sandboxes;
Intel macOS works in
static-only mode; native Windows is unsupported).
- Python: 3.12 – 3.14 with
venv (on Debian/Ubuntu:
sudo apt install python3-venv).
- LLM credentials (pick one):
- Vertex AI:
gcloud auth application-default login (with an active GCP
project or export GOOGLE_CLOUD_PROJECT=<project-id>)
- Gemini API:
export GEMINI_API_KEY=...
- Anthropic / OpenAI / custom OpenAI-compatible endpoints: export the
provider API key (
ANTHROPIC_API_KEY, OPENAI_API_KEY, etc. per
LiteLLM Providers) and select
a matching --model in Step 2 below
1. Install
cd reference && ./install.sh
install.sh creates reference/.venv and, if docker/podman/buildah is
available, builds the optional local sandbox image for dynamic reproduction. If
no container/VM tooling is available, Mantis falls back cleanly to
static-only.
2. Configure & verify
# Auto-detect sandbox capabilities and Vertex AI / GEMINI_API_KEY defaults
# (or pass --interactive for a guided wizard). Writes gitignored workflow.local.json.
.venv/bin/python3 scripts/configure.py --auto
# Non-Vertex providers: set --model explicitly (required for Anthropic, OpenAI, and --api-base)
# Gemini API: .venv/bin/python3 scripts/configure.py --model gemini-3.8-flash
# Anthropic: .venv/bin/python3 scripts/configure.py --model anthropic/claude-sonnet-4-5
# OpenAI/custom: .venv/bin/python3 scripts/configure.py --model openai/gpt-4o --api-base https://api.openai.com/v1
# Optional: force static-only mode (zero VM/container requirements)
.venv/bin/python3 scripts/configure.py --sandbox static-only
# Fast preflight check (~1s) + live LLM reachability probe
.venv/bin/python3 scripts/configure.py --test --probe
3. Run your first scan
./run.sh activates reference/.venv automatically and accepts either a single
file or a repository directory:
# Balanced full-stack + firmware/RTL scan (defaults to --depth 30)
./run.sh /path/to/repo
# Fast triage (Critical/High focus, single campaign on typical repos)
./run.sh /path/to/repo --depth 10
# Exhaustive multi-pass deep audit across the repository
./run.sh /path/to/repo --depth 100
# Scan a single file
./run.sh /path/to/file.py
--depth takes any integer from 1 (fast shallow triage) to 100
(exhaustive full-coverage multi-pass audit) and defaults to 30. It smoothly
scales risk-ranked file coverage, batch granularity, reasoning effort,
deterministic sink checklists, and pass count while always respecting existing
budget guardrails (--token-budget, --max-time) and explicit CLI overrides
(--scan-mode, --passes, --reasoning-effort, --model, --sandbox). See
reference/README.md for the full breakdown.
4. Inspect results & use Mantis during development
Every run prints a summary to the terminal and saves verified findings, threat
models, risk scores, coverage, and learnings in reference/knowledge.db (custom
path via --db):
# Query accumulated threat models, findings, and prior fixes for a file
.venv/bin/python3 scripts/advise.py --file /path/to/repo/src/auth.py
- SARIF 2.1.0 export: set
"sarif_output": "workspace/mantis.sarif" in
workflow.local.json (see
reference/README.md).
- IDE / coding-agent sidecar (MCP): connect MCP-capable coding agents (such
as Gemini CLI) via reference/docs/mcp.md for live
mantis_check_change diff gating and background scans.
Common Options
# One-shot model or sandbox override without changing workflow.local.json
./run.sh /path/to/repo --sandbox static-only --model vertex_ai/gemini-3.8-flash
# Steer the planner in plain language, or exclude vendored/generated paths
./run.sh /path/to/repo --focus "look for IDOR"
./run.sh /path/to/repo --exclude vendor,testdata --exclude 'third_party/*'
# Choose an explicit scan mode, or verify fixes incrementally after editing code
./run.sh /path/to/repo --scan-mode file-by-file # or whole, uncovered, cross-functional, auto
./run.sh /path/to/repo --scan-mode changed # new run after edits (omit --resume)
# Resume an interrupted or budget-paused run on the same snapshot
./run.sh /path/to/repo --resume # inherits recorded scan mode
# Export/import learned knowledge alongside a repository (.mantis/knowledge/)
.venv/bin/python3 scripts/knowledge.py export --db knowledge.db --repo /path/to/repo
./run.sh /path/to/repo --import-knowledge /path/to/repo/.mantis/knowledge
# Control spend ceilings and parallelism (SQLite lock wait defaults to max(60s, 15s x --parallel))
./run.sh /path/to/repo --max-tokens 2M --cache-discount 1
./run.sh /path/to/repo --no-budget --parallel 32 --db-timeout 600 # or MANTIS_DB_TIMEOUT=600
# Synthesize a custom multi-agent research graph for a specific objective
./run.sh /path/to/repo --objective "Audit for SSRF in webhook handlers"
See reference/README.md for the complete CLI reference,
scan modes, sandboxes, and Harness² evolution loop.
Overview of Mantis
Mantis is roughly designed to:
- Review history of a codebase to look for historical vulnerabilities we do not
wish to repeat
- Build a semantic tree-sitter index (20+ native grammars across backend,
frontend, mobile, smart contracts, shell/IaC/policy, and firmware/RTL) and
discover repository docs (
README.md, SECURITY.md, THREAT_MODEL.md,
architecture specs) as untrusted intent context to verify against real code
- Automatically generate a threat model
- Build up a set of hypotheses for individual agents to research (or simply
"scan every file")
- Execute on those research plans
- Deduplicate existing findings
- Triage/critique findings to combat hallucination and statically verify
production viability
- Reproduce the vulnerability to varying degrees, depending on available
environments, anywhere from a static guess at what an exploit might look like
up to a unit test or even spinning up a mock server to attempt to exploit
- Look across known vulnerabilities to attempt to build more impactful chains
- Patch discovered vulnerabilities, using an adversarial loop to verify the
vulnerability is really fixed
- Calibrate all findings based on an established rubric to combat LLM inflation
of severity (surfacing the most critical risks to humans instead of spewing
thousands of "criticals")
- Reflect on each iteration of the loop to look for things we've learned in the
now completed round of research
- Based on all of the collected learnings, threat model, and knowledge base, use
the
/mantis-advise skill (or the MCP server) to
develop code more securely and ensure that during development you do not
repeat prior mistakes or trigger edge cases in code that were previously
protected by some guard that was removed
Additionally, the ADK reference harness shows some of the neat ways in which we
can build agentic workflows. Specifically, reference/workflow.json shows how a
deep review is done, and research graph synthesis
(./run.sh target --objective "...") allows generating custom graph topologies
on the fly. This is a very powerful construct because you can use this to build
any kind of agentic deep dive security review you might imagine.
Mantis works out of the box across full-stack web apps, backend/systems code,
firmware/RTL (Verilog, SystemVerilog, VHDL, embedded/eBPF), Smart Contracts/ZK,
CI/CD supply chains, mobile (Swift, Objective-C, Dart, Android), AI/ML/agentic
pipelines, desktop/browser extensions, and Cloud/IaC/policy repos — with
blocklist-default file selection so novel languages and DSLs are never silently
dropped. It is also intended to be a starting point rather than a rigid set of
instructions: you can further adapt, tune, and extend the
skills and specialized domain guides
to fit your organization's specific stack. We provide an ADK-based reference
harness which works out of the box, but any competent coding agent should be
able to convert this to use your framework of choice.
We strongly recommend using AI to iterate on Mantis or integrate it with your
existing harness. Use your internal documentation, coding standards, and build
systems to augment the threat models you use for scanning. We also strongly
recommend adapting risk calibration to your environment and risk tolerance.
Above all, while orchestrated vulnerability discovery is incredibly powerful and
useful, it is even more important to use this in a suitably isolated environment
to prevent impacting production systems. The ADK reference harness provides some
example sandboxes for testing, but if you use the most advanced frontier models
you should go beyond this and additionally set up an additional sandboxing layer
that itself contains very strong monitoring to look for escape attempts.
License
Mantis is licensed under the Apache License, Version 2.0. Every
Python and shell file carries the license header;
reference/scripts/check_license_headers.py (run by pre-commit and the test
suite) keeps it that way.
Disclaimers
This is not an officially supported Google product. If you are looking for a
managed and hosted code security agent rather than building your own, we
recommend CodeMender. This
project is not eligible for the
Google Open Source Software Vulnerability Rewards Program.
This project is intended for demonstration purposes only. It is not intended for
use in a production environment.